Self-assembly on Demand in a Group of Physical Autonomous Mobile Robots Navigating Rough Terrain

Transcription

1 Selfassembly on Demand in a Group of Physical Autonomous Mobile Robots Navigating Rough Terrain Rehan O Grady 1,RoderichGroß 1, Francesco Mondada 2, Michael Bonani 2,andMarcoDorigo 1 1 IRIDIA, Université Libre de Bruxelles, Brussels, Belgium {rogrady, rgross, mdorigo}@ulb.ac.be 2 ASL, Ecole Polytechnique Fédérale de Lausanne, Lausanne, Switzerland {francesco.mondada, michael.bonani}@epfl.ch Abstract. Consider a group of autonomous, mobile robots with the ability to physically connect to one another (selfassemble). The group is said to exhibit functional selfassembly if the robots can choose to selfassemble in response to the demands of their task and environment [15]. We present the first robotic controller capable of functional selfassembly implemented on a real robotic platform. The task we consider requires a group of robots to navigate over an area of unknown terrain towards a target light source. If possible, the robots should navigate to the target independently. If, however, the terrain proves too difficult for a single robot, the robots should selfassemble into a larger group entity and collectively navigate to the target. We believe this to be one of the most complex tasks carried out to date by a team of physical autonomous robots. We present quantitative results confirming the efficacy of our controller. This puts our robotic system at the cutting edge of autonomous mobile multirobot research. 1 Introduction Collective robotics addresses the design, implementation and study of multirobotic systems. Swarm robotics is a subset of collective robotics which takes inspiration from social insect behaviour and emphasises swarm intelligence [2] principles such as decentralisation of control and use of local information. Many swarm robotics applications require cooperation between robots [8]. Some applications further require physical connectivity between cooperating robots. It is this last class of application that interests us. Although there is a large body of work on the capabilities of physically connected systems, very little research has been conducted on the mechanisms of when and how autonomous mobile agents should selfassemble. The phrase functional selfassembly [15] describes a key adaptive response mechanism of distributed systems. We define selfassembly as the process M. Capcarrere et al. (Eds.): ECAL 2005, LNAI 3630, pp , c SpringerVerlag Berlin Heidelberg 2005

2 Selfassembly on Demand in a Group (a) 273 (b) Fig. 1. (a) Sbots overcome a 2 cm hill independently. (b) Sbots selfassemble in order to overcome a 5 cm hill collectively. through which separate autonomous agents form a larger group entity by physically connecting to one another. If the agents can autonomously choose to selfassemble in response to the demands of their task and environment, they are said to display functional selfassembly. A number of social insect species depend on functional selfassembly (for a review see [1]). Members of the ant species Œcophylla longinoda, for example, connect to one another to form bridges that other ants can then traverse [7]. Given its ubiquity in natural systems, functional selfassembly has been given surprisingly little attention by the swarm robotics community. In the only dedicated work, Trianni et al. [15] evolved neural network controllers for robots that needed to selfassemble and disassemble in order to traverse artificially designated hot and cold zones in a simple simulation environment. Over the last decade, much of the research involving systems of physically connected robotic modules has been targeted at collective rough terrain navigation. In Hirose et al. s system [6] modules are mechanically linked by means of a passive arm and are therefore incapable of selfassembly. Yim et al. s system [16] can climb near vertical walls. Individual modules are incapable of autonomous motion and have very few external sensors for perception of the environment. Similar limitations are found in the majority of selfreconfigurable robotic systems, usually rendering selfassembly difficult or impossible [12,14]. In this paper we present the first physical robot controller capable of functional selfassembly. Our controller was implemented on the SWARMBOT robotic platform [11,10,3]. This innovative system consists of a number of autonomous robotic agents called sbots. Sbots are able to physically connect to one another, thus forming a larger group entity termed a swarmbot. A swarmbot can complete tasks impossible for a single sbot. It can, for example, cross chasms wider than an sbot or overcome hills too steep for a single sbot. The task we investigate requires a group of sbots to navigate towards a target light source over unknown terrain. The sbots must decide whether or not to selfassemble based on the terrain they encounter. We use two different environments in our experiments. The first environment contains a simple hill which a single sbot can overcome (see Fig. 1a). The sbots can thus reach the

3 274 R. O Grady et al. (a) (b) Fig. 2. (a) The sbot. (b)thesbot gripping mechanism. target independently. The second environment contains a steep hill too difficult for a single sbot. Thesbots must selfassemble in order to overcome the hill and reach the target (see Fig. 1b). 2 Experimental Setup 2.1 The SBot This study was conducted on the SWARMBOT robotic platform [11,10,3]. The system consists of a number of mobile autonomous robots called sbots (see Fig. 2a). The sbot is equipped with a traction system made up of tracks and wheels. This chassis provides the sbot with efficient on the spot rotation and mobility on moderately rough terrain. The majority of the sbot sensory and processing systems are housed in a turret mounted above the chassis. A motorised axis allows this turret to rotate with respect to the chassis. Physical connections between two sbots can be established by a gripperbased connection mechanism (see Fig. 2b). Each sbot is surrounded by a T shaped ring which can be grasped by other sbots. The sbot sensory systems used in this study are as follows: 15 proximity sensors distributed around the turret allow for the detection of obstacles. A 3 axes accelerometer provides information on the sbots inclination which can be used to detect if the sbot is in danger of falling. The connection ring of the sbot is equipped with eight groups of coloured LEDs. An omnidirectional camera is mounted on the turret. The combination of the camera and the LED ring allows an sbot to communicate its presence and even its internal state to other nearby sbots. Inside the gripper is an optical light barrier to detect the presence of objects to be grasped. Other sensors provide the sbot with information about its internal motors. This includes positional information (e.g., of the rotating turret) and torque information (e.g., of forces acting on the tracks). 2.2 The Task We conduct experiments in two different environments (see Fig. 3). Both measure 240 cm x 120 cm and consist of two areas of flat terrain (a starting area and a target area) separated by an area of rough terrain. In Environment A, the rough

4 Selfassembly on Demand in a Group 275 Starting Area Rough Terrain Environment A Flat Terrain Target Area Light Source Starting Area Target Area Rough Terrain Light Source Environment B 01 Simple Hill Single s bot succeeds alone 01 Difficult Hill Single s bot fails Fig. 3. Scale diagram of the two experimental environments (view from above). Sbot starting positions are marked by crosses. terrain is a 2 cm high hill which can be overcome by a single sbot (see Fig. 1a). In Environment B the rough terrain hill is 5 cm high too difficult for a single sbot (see Fig. 1b). The initial position of each sbot in the starting area is assigned randomly by uniformly sampling without replacement from a set of 15 possible starting points. The sbot s initial orientation is chosen randomly from a set of 4 possible directions. To complete the task the sbots must reach the target area without toppling over. The sbots have no a priori knowledge of the environment they are in they must react to the environment and determine whether or not to selfassemble. In Environment A the sbots should navigate to the target area independently. In Environment B the sbots must aggregate, selfassemble and collectively overcome the hill in order to reach the target area. 3 Controller We use a distributed behaviourbased controller (see Fig. 4). Each sbot is fully autonomous. The same controller is executed on every sbot. An sbot starts by navigating independently towards the target light source. If the sbot finds a hill toodifficultforittopassalone,orifitseesanothersbot that is either aggregating or assembled (sees blue or red), it illuminates its blue LEDs and starts aggregating. An aggregating sbot can probabilistically trigger selfassembly by Prob(Become Seed) and Close to Blue and Don t See Red Assembly_Seed (red) Timeout Over Solo_Phototaxis Avoid_Obstacle See Blue or See Red Too Steep Aggregate (blue) See Red Within Timeout On Flat Retreat_to_Flat (blue) Close to Red Self_Assembly Group_Phototaxis (red) (See blue => Wait) Assembled Fig. 4. Behaviour transition model for the behaviourbased controller

5 276 R. O Grady et al. Aggregate behaviour Self Assembly behaviour 1: loop 2: if canseeclose( red ) then 3: switchbehaviour( Self Assembly ) 4: else if canseefar( red ) then 5: approachred( ) 6: else if canseeclose( blue ) then 7: Prob(0.04) switchbehaviour( Assembly Seed ) 8: Prob(0.96) donothing( ) 9: else if canseefar( Blue ) then 10: approachblue( ) 11: else 12: randomwalk( ) 13: end if 14: end loop 1: repeat 2: (i 1,i 2) featureextraction(camera) 3: (i 3,i 4) sensorreadings(proximity) 4: (o 1,o 2,o 3) f(i 1,i 2,i 3,i 4) 5: if graspingrequirementsmet(o 3) then 6: try to grasp 7: else 8: applyvaluestotracks( o 1, o 2 ) 9: end if 10: until successfully connected Fig. 5. Algorithms for Aggregate behaviour (left) and Self Assembly behaviour (right) illuminating its red LEDs and becoming a static seed. Aggregating sbots assemble to the seed sbot or to already assembled sbots (any red object). Assembled sbots illuminate their red LEDs then perform group phototaxis once they can no longer detect any unassembled sbots (can no longer see blue). Solo Phototaxis. This is the starting behaviour. The sbot uses its camera to navigate towards the target light source. The sbot uses its accelerometers to reduce maximum track speed as a linear function of inclination. This is to prevent the sbot toppling before Retreat to Flat behaviour is triggered. Avoid Obstacle. This behaviour is triggered when the readings from the sbot s 15 proximity sensors exceed a threshold. The sbot determines the direction of the obstacle using its proximity sensors then moves in the opposite direction until the proximity threshold is no longer exceeded. Retreat to Flat. This behaviour is initiated when the sbot s accelerometers indicate that the sbot is in danger of toppling over. The sbot reverses downhill to flat terrain, reverses away from the rough terrain, then rotates to face away from the slope. Aggregate. This behaviour is detailed in Fig. 5 (left). The sbots must locate and then approach each other as a precondition for selfassembly. Values for the hard coded probabilities were manually optimised through trial and error. Self Assembly. This behaviour is detailed in Fig. 5 (right). Function f maps sensory input (i 1,i 2,i 3,i 4 ) to motor commands (o 1,o 2,o 3 ). It is implemented by a neural network which was designed by artificial evolution and tested with physical robots in previous works [5,4]. Assembly Seed. This behaviour is necessary to trigger the selfassembly process. If a red object is detected within 3 s of behaviour initiation, control is passed to Aggregate behaviour. (This prevents multiple seeding if two nearby sbots switch to Assembly Seed behaviour, both will revert to Aggregate behaviour). After 3 s control is passed to Group Phototaxis behaviour.

6 Selfassembly on Demand in a Group 277 Group Phototaxis.The sbot remains stationary if it detects blue objects in the vicinity (sbots still assembling). Otherwise the sbot performs phototaxis to the target. Because it is part of a swarmbot, the orientation of the turret is fixed. The sbot continually rotates the traction system with respect to the turret to keep the tracks oriented towards the target [5]. 4 Results We conducted a series of experiments in two different environments (see Fig. 3) with groups of 1, 2 and 3 sbots. 1 Trials with 3 sbots in Environment A. We conducted 20 trials. In every trial all 3 sbots reached the target zone. In 19 out the 20 trials the sbots correctly navigated independently to the target. In a single trial the sbots selfassembled on the down slope of the hill and then performed collective phototaxis to the target. The incorrect decision to selfassemble was due to a colour misperception of a nonexistent object by an sbot. Trials with a single sbot in Environment B. We modified the controller to only execute Solo Phototaxis behaviour. The sbot was thus limited to navigating towards the target taking no account of the terrain encountered. We conducted 20 trials. The sbot failed to overcome the hill in 20 out of 20 trials. In each trial the sbot reached the hill and then toppled backwards due to the steepness of the slope. To confirm that the sbot was failing due to the intrinsic properties of the slope, we repeated this experiment at a number of different constant speeds. Trials with 2 sbots in Environment B. We conducted 20 trials. The s bots successfully detected the slope in every trial. Furthermore the sbots always succeeded in assembling into a 2 sbot swarmbot. In 13 trials (65%) the swarmbot succeeded in overcoming the hill. In the other 7 trials (35%) the assembled swarmbot failed to overcome the hill. These failures happened when the assembled sbots attempted to climb the hill in parallel. Trials with 3 sbots in Environment B. We conducted 20 trials. The sbots successfully detected the slope in every trial. In 16 trials (80%) all of the sbots successfully selfassembled into a 3 sbot swarmbot. In each of these 16 trials the 3 sbot swarmbot went on to successfully reach the target area. Fig. 6 shows a sequence of images from a typical trial. In the remaining 4 trials (20%) the sbots still managed in each case to selfassemble into a swarmbot of 2 sbots. In two of these 4 trials the swarmbot went on to successfully reach the target area. In the two other trials the swarmbot was obstructed by the third sbot which failed to selfassemble. 1 Videos of all experiments can be found at rogrady/ ecal2005/

7 278 R. O Grady et al. (a) (b) (c) (d) (e) (f) Fig. 6. The sbots start in a random configuration (a). One sbot detects a slope it cannot overcome alone and activates blue LEDs (b). The other sbots detect blue colour (local communication). The group aggregates and selfassembles (c,d). The sbots collectively overcome the rough terrain and reach the target area (e,f). Table 1. Percentage of sbots in Environment B trials succeeding for Selfassembly (A) and Completion of task (C) 1 sbot trials 2 sbot trials 3 sbot trials A 4.1 C A C A C % Successful (total) % Successful alone 0.00 % Successful in 2 sbot swarmbot % Successful in 3 sbot swarmbot % Failed Analysis Table 1 shows the percentage of sbots that successfully selfassembled and the percentage of sbots that successfully completed the entire task in the Environment B experiments. The three columns distinguish between trials with 1 sbot, 2 sbots and 3 sbots. The first row shows the total percentage of successful sbots. Subsequent rows show the percentage of sbots that were successful alone, or as part of a 2 sbot swarmbot or as as part of a 3 sbot swarmbot, or that failed. The success rate for task completion increases with the number of robots. A single robot always fails. In 2 sbot trials, 65% of sbots complete the task. The 3 sbot trials show a further clear improvement 86.67% complete the task.

8 Selfassembly on Demand in a Group 279 Fig sbot trials in Environment B. Phases represented are: (i) sbot: independent s bot navigation; (ii) assembly: aggregation and selfassembly; (iii) swarmbot: collective swarmbot navigation. The fourth row (% Successful in 3 sbot swarmbot) shows that in the 3sbot trials 80% of sbots successfully selfassemble into a 3 sbot swarmbot. Thesame row shows us that 80% of sbots complete the task in a 3 sbot swarmbot. Thus in 3 sbot trials, whenever all the 3 sbots successfully selfassemble into a 3 s bot swarmbot they always successfully overcome the rough terrain. By contrast, in the 2 sbot trials 100% of the sbots selfassemble into a 2 sbot swarmbot. Despite this, only 65% of the 2 sbot swarmbots successfully overcome the hill. The hill in environment B is such that in our trials a 3 sbot swarmbot always (100% of the trials) overcomes it. A 2 sbot swarmbot on the other hand sometimes (35% of the trials) fails to overcome the hill. Whenever the 2 sbot swarmbot approached the hill in parallel the swarmbot toppled backwards. Fig. 7 illustrates three phases of task completion. In the first phase (black segment) all sbots are independently navigating to the target (this phase ends when the hill is first detected by an sbot). The phase takes between 4 s and 17 s depending on the random initial configuration of the sbots. For the unsuccessful trials (4,8,12,16) only this first phase is illustrated. The second phase (white segment) consists of aggregation and selfassembly. This phase takes between 39 s and 175 s. This phase always accounts for a large percentage of total completion time due to its high level of complexity. The final phase (grey segment) consists of collective phototaxis to the target. This phase takes between 4 s and 30 s, except in trial 17, when the swarmbot got stuck for some time on the hill. Thesymbol c infig.7marksthefirsttimethatallsbots become aware of the hill. In some trials (e.g. trials 5 and 6) the existence of the difficult hill is communicated very quickly between sbots (see also Fig. 6). One sbot detects the rough terrain and activates its blue ring LEDs. The other sbots are already close enough to detect this blue colour. In such trials the point c is reached soon after the start of the aggregation and selfassembly phase. In other trials

9 280 R. O Grady et al. (e.g. trials 1 and 12) it takes longer to reach point c as the sbots are sufficiently far apart that two sbots discover the hill independently. The symbol s in Fig. 7 indicates when selfassembly was seeded (the last time an sbot switches to Assembly Seed behaviour). 5 Conclusion Selfassembly is a critical adaptive response mechanism in a number of social insect species. This work represents the first successful use of this response mechanism by real robots. We have shown that a group of physical autonomous mobile robots can choose to selfassemble in response to the demands of their task and environment. Using our controller, a group of robots faced with a simple hill overcome it independently. When the same robots are faced with a hill too difficult for a single robot they selfassemble and overcome the hill together. The success rate increased with the number of robots used: 0%,65% and 86.67% for groups of 1, 2 and 3 robots respectively. Our approach involved splitting the task (as seen from the perspective of an individual robot) into distinct phases. Each phase was addressed by a separate behaviour module these modules were combined to produce our behaviour based controller. In a previous work conducted in a simplified simulation environment, Trianni et al. [15] focused on evolving a single neural network controller to achieve functional selfassembly. We believe that application of this evolutionary approach to the real robots might yield solutions that exploit hidden properties of the robotic hardware or which make better use of the complex group dynamics of the task [13]. We are currently investigating mechanisms to generate connection patterns and group sizes [9] that are suited to particular tasks. In the spirit of functional selfassembly we would like the robots themselves to choose these patterns and group sizes as they interact with their environment. Acknowledgements This work was supported by the ANTS project, an Action de Recherche Concertée funded by the Scientific Research Directorate of the French Community of Belgium; by the SWARMBOTS project, funded by the Future and Emerging Technologies programme of the European Commission (grant IST ); and by COMP2SYS, a Marie Curie Early Stage Research Training Site funded by the European Community s Sixth Framework Programme (grant MEST CT ). The information provided is the sole responsibility of the authors. It does not reflect the Community s opinion. The Community is not responsible for any use that might be made of data in this publication. Marco Dorigo acknowledges support from the Belgian FNRS, of which he is a Research Director.

10 Selfassembly on Demand in a Group 281 References 1. C. Anderson, G. Theraulaz, and J.L. Deneubourg. Selfassemblages in insects societies. Insectes Soc., 49:99 110, E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence: From Natural to Artificial Systems. Oxford University Press, New York, NY, M. Dorigo, V. Trianni, E. Şahin, R. Groß, T. H. Labella, G. Baldassarre, S. Nolfi, J.L. Deneubourg, F. Mondada, D. Floreano, and L. M. Gambardella. Evolving selforganizing behaviors for a Swarmbot. Auton. Robots, 17(2 3): , R. Groß, M. Bonani, F. Mondada, and M. Dorigo. Autonomous selfassembly in mobile robotics. Technical Report IRIDIA/20052, IRIDIA Université Librede Bruxelles, Submitted to IEEE Trans. Robot. 5. R. Groß and M. Dorigo. Group transport of an object to a target that only some group members may sense. In Parallel Problem Solving from Nature 8th Int. Conf. (PPSN VIII), volume 3242 of Lecture Notes in Computer Science, pages Springer Verlag, Berlin, Germany, S. Hirose, T. Shirasu, and E. F. Fukushima. Proposal for cooperative robot Gunryu composed of autonomous segments. Robot. Auton. Syst., 17: , A. Lioni, C. Sauwens, G. Theraulaz, and J.L. Deneubourg. Chain formation in Œcophylla longinoda. J. Insect Behav., 15: , A. Martinoli, K. Easton, and W. Agassounon. Modeling swarm robotic systems: A case study in collaborative distributed manipulation. Int. J. Robot. Res., 23(4): , C. Melhuish, O. Holland, and S. Hoddell. Convoying: Using chorusing to form travelling groups of minimal agents. Robot. Auton. Syst., 28: , F. Mondada, L. M. Gambardella, D. Floreano, S. Nolfi, J.L. Deneubourg, and M. Dorigo. SWARMBOTS: Physical interactions in collective robotics. IEEE Robot. Autom. Mag., 2005, to appear. 11. F. Mondada, G. C. Pettinaro, A. Guignard, I. V. Kwee, D. Floreano, J.L. Deneubourg, S. Nolfi, L. M. Gambardella, and M. Dorigo. SWARMBOT: A new distributed robotic concept. Auton. Robots, 17(2 3): , S. Murata, E. Yoshida, A. Kamimura, H. Kurokawa, K. Tomita, and S. Kokaji. M TRAN: Selfreconfigurable modular robotic system. IEEE/ASME Trans. Mechatron., 7(4): , S. Nolfi and D. Floreano. Evolutionary Robotics: The Biology, Intelligence, and Technology of SelfOrganizing Machines. MIT Press, Cambridge, MA, M. Rubenstein, K. Payne, P. Will, and W. M. Shen. Docking among independent and autonomous CONRO selfreconfigurable robots. In Proc. of the 2004 IEEE Int. Conf. on Robotics and Automation, volume 3, pages IEEE Computer Society Press, Los Alamitos, CA, V. Trianni, E. Tuci, and M. Dorigo. Evolving functional selfassembling in a swarm of autonomous robots. In From Animals to Animats V III.Proc.ofthe8 th Inter. Conf. on Simulation of Adaptive Behavior, pages MIT Press, Cambridge, MA, M. Yim, K. Roufas, D. Duff, Y. Zhang, C. Eldershaw, and S. B. Homans. Modular reconfigurable robots in space applications. Auton. Robots, 14(23): , 2003.